We have introduced a new approach to adaptive prediction-error
filtering for seismic data interpolation. Our approach uses
regularized nonstationary autoregression to handle time-space
variation of nonstationary seismic data. We apply this method to
interpolating seismic traces beyond aliasing and to reconstructing
data with missing and decimated traces. Experiments with benchmark
synthetic examples and field data tests show that the proposed filters
can depict nonstationary signal variation and provide a useful
description of complex wavefields having multiple curved events. These
properties are useful for applications such as seismic data
interpolation and regularization. Other possible applications may
include seismic noise attenuation.